The MVSep Bass model exists in 3 different variants based on following architectures: HTDemucs4, BS Roformer and SCNet XL. The model produces high-quality separation of music into a bass part and everything else.
Quality metrics
| Algorithm name | Multisong dataset | MDX23 Leaderboard |
|
| SDR Bass | SDR Other | SDR Bass | |
| BS Roformer | 12.49 | 16.59 | --- |
| HTDemucs4 | 12.52 | 16.64 | --- |
| SCNet XL | 13.81 | 17.93 | --- |
| BS + HTDemucs + SCNet XL Ensemble | 14.07 | 18.18 | --- |
| BS + HTDemucs + SCNet XL Ensemble (+extract from Instrumental) | 14.12 | --- | --- |
| BS Roformer SW | 14.62 | --- | --- |
| SCNet XL + BS Roformer SW | 14.87 | --- | --- |
